PENTA-CL Interim Results: Canonical Replay Matches and Step-String Drift in Four Platform-Configured LLM Engines

Abstract

Audited interim result, not completed PENTA certification. Four platform-configured engines were observed in two fresh web conversations each with memory disabled, using a frozen six-case synthetic decision prompt. Opus 5.5 and GPT-6.1 Sol matched the registered canonical JSON tuple hash in both recorded captures. GPT-5.6 Luna matched in A but differed in B. GLM 5.2 replayed consistently but differed from the registered target in both captures. DeepSeek has not been captured. All 48 recorded decisions selected the expected playbook name and id. Full registered tuple equality held for 30/48 decisions. The remaining 18 differed only in spaces inside the steps_executed string. Pretty-printing outside strings is removed by canonicalization and is not a failure cause. The prompt's lack of an explicit space delimiter contract is disclosed; failures remain failures against the preregistered target. Limitations: UI-transcribed rendered captures, not raw provider responses; model attribution by Base44 configuration, not backend attestation; supplementary fresh-chat method deviates from original same-conversation registration; shared instructions, workspace and expected answers remain available; Sol executed skills in both captures; Opus A conversation attribution was inferred from sidebar/timestamp ordering. This is not a stress/load test, live-healing test, general governance certification, or proof of runtime enforcement. Confidence-hold and no-match branches were not exercised. Audit and correction: earlier copied Luna files and single-turn Opus/Sol attempts are excluded, preserved, and labeled. Earlier operator claims of raw byte-exactness and 100% governance are corrected explicitly. Deterministic Phase A/C receipts are harness-computed, not LLM inference evidence. No failure is normalized into a pass. Dual Mesh Benchmark: not measured in this experiment. No comparative mesh, latency, learning or healing performance claim is made. Bundle contains the technical note in PDF/Markdown, machine-readable receipt, full evidence archive, offline verifier, original preregistration, excluded attempts, and SHA-256 manifest. All test data is synthetic. AI assistance and capture provenance are disclosed. Registration: public commit deadfea. Results: GitHub evidence. Research home: DCAI; portal: DSOS Science. Lineage acknowledged: Nyquist, Hartley, Shannon, Kolmogorov. No new mechanism is claimed. This software is a prototype and is provided for educational and research purposes only. It is not intended for production use, commercial deployment, or safety-critical environments. All systems are experimental and may contain defects on them.